Accuracy benchmark · octubre 2026

98% of numbers correct across 82 hand-checked real-world pages

We run our production extraction pipeline on real documents and compare every number, word and date with text checked by hand against the source. For numbers: 99% on the 50 pages held out from tuning.

Numbers correct, all 82 pages
98,1%
Numbers correct, 50 pages held out from tuning
99,1%
Bank-statement rows flagged by the balance check
1 / 284

Results

“Correct” is the share of values in our output that appear exactly on the page. “Found” is the share of values on the page that appear in our output.

Páginas Numbers correct Numbers found Words correct Words found Dates correct Dates found
Todas las páginas (82) 98,1% 97,3% 96,2% 90,2% 99,5% 99,5%
Held out from tuning (50) 99,1% 98,4% 96,6% 93,0% 99,7% 100%
Numbers by kind of page
Páginas Correct Found
Digital PDFs (29) 99,7% 99,4%
Escaneos (41) 97,2% 96,8%
PDFs with a scanner's text layer (9) 98,2% 98,6%
Fotos del teléfono (3) 96,4% 92,3%
Numbers by kind of document
Páginas Correct Found
Financial tables (statements, invoices) (32) 98,7% 97,5%
Plain text documents (11) 92,9% 94,3%

Other layouts are not broken out here; every page counts in the overall results.

Extractos bancarios

Of 284 transaction rows extracted from the statements in the set, the running-balance check flagged 1 for review. When a statement prints its balances, the preview also checks that opening balance + credits − debits = closing balance, and tells you when they match.

How we measure

  • 68 real-world documents (82 pages): bank statements, invoices, receipts, forms, scans and phone photos, sampled to match the mix of files people upload.
  • Every page's expected text was checked by hand against the source document.
  • A number counts as correct only if the exact value appears on the page, and as found if a number on the page appears in our output. Words and dates are scored the same way.
  • 50 pages were held out and never used while tuning the pipeline.
  • Results are from the production pipeline, scored on octubre 2, 2026.

Limitations

  • The set has only 3 phone photos, so the photo figures rest on a small sample.
  • Plain text documents score lower than tables: 92,9% of numbers correct.
  • Results on your documents may differ. Check the preview against the original before you rely on it.
  • We re-run the benchmark when the pipeline changes and update this page.

Check it on your own document

Upload a PDF, scan or photo and compare the preview with the original.

Try it on a document